Estimating mean total costs in the presence of censoring

Abstract

Introduction: Frequently, within economic evaluations, data are subject to censoring, and ignoring censored data will lead to an underestimation of mean total costs. Several techniques have been published that can be used to estimate mean total costs and standard errors, and allow for censoring within cost data. This paper reviews these techniques and compares the mean total costs estimates generated by each method for different types of censoring.

Methods: Nine techniques were identified that can be used to estimate mean total costs and standard errors in the presence of censoring: ignoring censoring; ignoring censored costs; Lin’s method — with and without cost histories; weighted cost method — with and without cost histories; Lin’s regression method — with and without cost histories; and Carides’ regression method. These methods are compared across four different censoring mechanisms — random censoring, end-of-study censoring, informative censoring and partial censoring — by simulating the censoring mechanisms from a complete cohort of patients included in the CELT (Cost Effectiveness of Liver Transplantation) study.

Results: The observed mean cost and standard error from the CELT data were £36 045 and £1517 (1998 values). Estimates under informative censoring were the least accurate predictors of mean total costs and tended to overestimate mean costs by >£1000. Carides’ regression method predicted mean total costs to within £3 of the observed mean and represented one of the three most accurate methods for predicting mean total costs (together with the weighted cost method with known cost histories and Lin’s method with unknown cost histories). Lin’s method with known cost histories gave the least accurate estimates of mean total costs and underpredicted costs by £2137–4859 across censoring mechanisms. Carides’ method did not predict uncertainty around the mean costs well, and the weighted cost method with known cost histories and ignoring censoring were the best methods to use for estimating the standard error of the mean cost.

Conclusions: Further work should be carried out on other datasets to confirm the generalisability of these results. Although Carides’ regression method and Lin’s method with unknown cost histories were the best estimators of mean total costs across censoring mechanisms, the weighted cost method with known cost histories is the preferred method for obtaining an accurate estimate of the mean total cost alone and the uncertainty surrounding it; therefore, it should be used to estimate mean costs and standard errors when patient cost histories are known.